Our answers on privacy
A care organisation asked how Inora handles its data. This is my answer of 1 October 2025: no training on conversations, no name attached to a message, storage in Europe, and a library that grows only after experts check it. What has changed since is dated beneath it.
Written by
Founder and CEO
Essay8 min read
The five questions a care organisation put to us about privacy, transparency and the responsible use of AI in care were exactly the questions we had asked ourselves when we started on Inora. Our contact there had forwarded them, and I read them with a smile. By then Inora had been working in the first care organisation for almost four weeks.
At the end of my answer I wrote that many organisations rightly ask these questions, and that the answers would go on our website. Here they are, shortened, in my words of the time, translated from the Dutch. I leave out the names of the organisation and of the people who asked.
Where an answer has changed since, the change is set out at the end, with its date. An answer on privacy that never changes is usually an answer nobody keeps up.
Is anyone training on the conversations?
That was their core question, so I started with it, and the answer was short: no.
We used the language models in Inora through their makers’ business access, under contracts that state our data is not used to train their models. We train nothing ourselves. We could not if we wanted to: we call a model, we do not retrain it.
In that email I drew a distinction I still think matters, because it gets muddled in almost every conversation about AI. Training a model on conversations is one thing. People reading conversations to make a product better is another.
We did not do the first. We did do the second. Our team read a selection of conversations by hand. If someone found an answer about diabetes incomplete and clicked the thumbs-down, we read that message. We saw no name, only a technical number. Our conclusion might be that the sources on diabetes needed extending. The model learned nothing from it. The library did.
Our team has since stopped that reading. The reason is at the end.
What we store
We stored the messages, with timestamps and session numbers that keep a conversation together. No name with a message, no email address with a conversation. I gave an example: if I look in the database, I see that someone asked about diabetes on Monday 15 September at 14:30, but not who. That link did not technically exist in our system.
The data sat in a closed-off environment in Europe, and only the developers who built Inora could reach it. We kept it for one purpose: to see where the library fell short, which questions kept coming back and where users got stuck.
Is the model frozen?
The fifth question was a sharp one, and the answer depended on what you call “the model”.
The language model itself: yes, frozen. It has finished learning before we use it, and it learns nothing from a conversation. The knowledge library Inora draws on: no, we keep that up to date. If many questions come in about a condition, we add sources. When a new guideline appears, it goes in. Such changes happened only after consultation and review by experts, never automatically.
I still use the comparison I made then. The language model works like a nurse’s mind once their training is complete: in a fixed way. The knowledge library is the reference shelf on the ward, updated again and again with new protocols and guidelines. The first stays the same; we keep the second current.
The librarian
I also sent the gist of the terms of use. Inora is an aid and does not replace professional judgement. Users remain responsible for their own decisions. And however reliable the sources, users always check an answer themselves and stay critical. You can only check if you can open the book. That is why an answer from Inora came with its source, and why Inora said so when the answer was not there.
At the end I added a question of my own, because I suspected all five were really about it: why is this different from a chat service anyone can use? My answer was a comparison. The difference is that between “an open public library where anyone can read and add every kind of book” and “a specialised medical library with only accredited and verified sources, where a librarian keeps a close watch on what comes in and who has access.”
For me that sentence still carries the whole answer. With AI in care, the privacy question is almost always about the model: what it remembers, who trains it, where it runs. Those are good questions, and the answers are above. But the questions that matter are about the library. What goes in, and who checked it? Who may enter? Who sees what someone else reads? A model that remembers nothing, in a library without a librarian, is still an open public library.
What has changed since
In 2025 I wrote that these answers described the state of affairs at that moment, and that we would build better privacy and security systems at a fast pace. This is what changed.
One answer did not change. No one trains a model on an organisation’s documents, questions, answers or reports, conversations included: not us, and not any model provider.
November 2025, client data. The terms of 2025 said staff enter no client data. That rule does not hold in a system where staff also write reports, because a report is about a client. We replaced it with what is true. The data processing agreement applies from first use and treats whatever a member of staff writes about a client as health data, from the first question. Inora never asks for a name, date of birth or citizen service number. A report needs none of them.
February 2026, who gets in. Staff sign in with the work account they already have, through their own organisation’s single sign-on, and lose access the day the organisation switches their account off. Who has access is the organisation’s decision. It also means a conversation now belongs to a named account, so that a member of staff can find their own conversations again. The separation I described in the database in 2025 is therefore now laid down in rules, which follow.
August 2026, who reads along. We no longer read a selection of conversations. To see where the sources fall short, a language model labels every message: answered or not, whether knowledge was missing, and what kind of knowledge. No one has to read a conversation for that. No one in the organisation reads a conversation, managers and the board included. Usage reports cover teams of at least five people and never name an employee, and Inora assesses no one. We read a conversation only when the organisation asks in writing, and we log every access.
September 2026, where and for how long. In 2025 I wrote where the data was stored. That was an answer about storage, and the better question is where the processing happens. All processing for Inora stays in the European Union: storage, search index, language models, sign-in and logging. Inora deletes conversations after 90 days, or after the period the organisation sets. And what I then called our core team of developers is now defined precisely: only the engineers who build Inora reach the production environment, each with a personal account and multi-factor authentication, never a shared account, and every access is logged.
September 2026, the library. Never automatically: that still holds for what we add. For an organisation’s own documents it works differently, and better. Every source in Inora has an owner, a version, a review date and a status, kept in step with the organisation’s own document system. When the organisation puts a new version in that system, Inora answers from it within minutes, without anyone at Axiomatic having to act. Inora stops using an expired or withdrawn document. The check sits where it belongs: with the organisation that wrote the document.
The topic of a question the sources could not answer lands as a task with the owner of that area of knowledge, with the number of times it was asked and never the question itself or the name of the person who asked it.
And the frozen model is replaced by a newer one now and then. We test every new version of Inora, a change of language model included, on the questions of the organisation’s own teams before the organisation receives that version.
In 2025 the librarian in my comparison was mostly a promise: we are paying attention. Now it is a set of rules, in the product, in the data processing agreement and in this piece, each with a date. That is what an answer on privacy should look like: not a sentence that reassures, but a list you can check.
Postscript, 2 November 2026
What a care worker asks Inora never reaches the organisation. Only what the care worker sends reaches it, such as a report or an incident report, and the competence record the inspectorate expects (Wkkgz, Wet BIG). That record (demonstrably competent, reserved procedures included) holds only that a protocol was followed, with its version and date, without a usage figure, a score or the content of a conversation.